Electronic copy available at: http://ssrn.

com/abstract=1312664
The Black-Litterman Model Explained

Wing CHEUNG

February, 2009
Abstract
Active portfolio management is about leveraging forecasts. The Black and Litterman Global Portfolio
Optimisation Model (BL) (Black and Litterman, 1992) sets forecast in a Bayesian analytic framework. In
this framework, portfolio manager (PM) needs only produce views and the model translates the views into
security return forecasts. As a portfolio construction tool, the BL model is appealing both in theory and in
practice.
Although there has been no shortage of literature exploring it, the model still appears somehow mys-
terious and suffers from practical issues. This paper is dedicated to enabling better understanding of the
model itself. It is featured by: -
An economic interpretation
A clarification of the model assumptions and formulation
An implementation guidance
A dimension-reduction technique to enable large portfolio applications
A full proof of the main result in the appendix
We also form a checklist of other practical issues that we aim to address in our forthcoming papers.
JEL Classification: C10, C11, C61, G11, G14
Keywords: asset allocation, portfolio construction, Bayes Rule, view blending and shrinkage, CAPM,
semi-strong market efficiency, mean-variance optimisation, robustness

This paper reproduces an earlier Nomura publication Cheung (2009b) with revisions. I would like to thank Gregory Bronner,
Ronny Feiereisen, Alan Hofmeyr, Hon Wai Lai, Wuan Luo, Amit Manwani, Attilio Meucci, Dushant Sharma, Stephen Vandermark,
and numerous European and US quant fund managers for discussions.

Corresponding author: Nomura International plc, 25 Bank Street, London E14 5LE, UK. Email: wing.cheung@nomura.com;
W.Cheung.02@cantab.net
1
Electronic copy available at: http://ssrn.com/abstract=1312664
1 Introduction
Active portfolio management is about leveraging forecasts. As a means of forecast, portfolio managers (PM)
or analysts collect information, generate views, and seek to convert them into optimal portfolio holdings.
These views may not necessarily be explicit security return predictions, but could be views on relative
performance or portfolio strategies
1
. On the other hand, portfolio optimisers do not admit views directly as
inputs, but rather expects one explicit return forecast for each security. In order to feed an optimiser, PMs
need to translate their views into explicit return forecasts for those view-relevant securities, and are forced
to come up with some number (often zero) to represent no view. This practice immediately attracts two
questions:
(1) What is the appropriate way of translating PM views into explicit return forecasts?
(2) Is it legitimate to use zero return to represent no view?
Regarding the second question, zero-mean return forecasts will be treated by the optimiser relentlessly
as views. A typical response of the optimiser will be to use this security to leverage others on which the
PM expresses optimism. This easily gives rise to unexpected behaviours (i.e., unstable, counter-intuitive
or corner solutions). Yet in this situation, imposing constraints is not the ultimate solution since this does
not address its underlying cause.
To the first question, the Black-Litterman Global Portfolio Optimisation Model (BL) (Black and Litter-
man, 1992) provides an elegant answer. The model sets forecast in a Bayesian analytic framework. In this
framework, the PM needs only produce a flexible number of views and the model smoothly translates the
views into explicit security return forecasts together with an updated covariance matrix - exactly as what a
conventional portfolio optimiser expects. If the views arrive in an acceptable form, i.e., linear views, this
model can fully consume them.
Moreover, the model handles the second question with ease: without view, there are theoretical justifications
for taking the market equilibrium returns as the default forecasts. A remarkable feature of this approach is
robustness. Since the posterior views are a combination of the market and the PM views, PMs have a com-
mon layer, the market view, as their starting point. Without views, the best strategy is to stick to the market
view. With some views, the portfolio should be tilted to reflect these views combined. Since the market view
is always considered, it is less likely to run into unstable or corner solutions. In case the PM holds some
strong views that dominate the market view, the model also allows the results to be significantly adjusted
towards these views. Rather than erratic, this should be considered as expected and intuitive.
Based on these, the BL model is appealing in theory and natural in practice. However, we still have not
seen wide applications of it. We attribute this to two main reasons:
The model deserves further explanation.
1
Views are often also expressed in terms of fundamental or macroeconomic factors. In a forthcoming paper, we will develop a
technique for factor-based allocation.
2
Electronic copy available at: http://ssrn.com/abstract=1312664
Some practical issues may have also frustrated applications, e.g., confidence parameter setting, alter-
native views (e.g., factor views, stock-specific views), curse of dimensionality (when applied to large
portfolios), prior setting (e.g., equilibrium is an abstract concept), optimiser issues, risk model quality,
non-linearity, non-normality issues etc.
Although there has been no shortage of literature exploring either the applications or the frontiers of
the model (e.g., Jones, Lim and Zangari, 2007; Meucci, 2006; Martellini and Ziemann, 2007; Zhou, 2008
etc.), we have seen few documents explaining it (see for example, Satchell and Scowcroft, 2000; Idzorek,
2004); let alone taking these practical issues seriously. We have therefore done some significant work to
make the model practical. As a starting point, this paper focuses on an explanation of the original model.
By exploring the information processing challenges encountered in a typical portfolio management process,
we enrich Black and Litterman (1992)s original motivation for the BL model. We then establish that the
BL model is buttressed by three pillars: the Semi-Strong Market Efficiency assumption, the Capital Asset
Pricing Model (CAPM), and the Bayes Rule. With the assistance of a carefully chosen notation system, we
formulate the model with particular attention to its technical details, i.e., model assumptions, main results,
and a full proof (in the appendix), to unveil the intrinsic logic. Implementation guidance then follows. In
order to enable large portfolio applications, we also discuss a dimension-reduction technique that resolves
the high-dimensionality issue before we reach our concluding remarks.
We will address a list of other practical issues in our forthcoming pieces as we develop new thoughts
around the topic.
Let us examine what motivates the BL model first.
2 Information Processing, Traditional Portfolio Optimisation and Its
Weaknesses
Suppose there are n securities in the investment universe. Assuming normality, the distribution of the se-
curity returns are fully determined by their first- and second moments, i.e.,

r
[n1]
N( m
[n1]
, V
[nn]
)
23
,
where m is the vector of real mean security returns and V is the real variance-covariance matrix. These
moments are not directly observable.
In a typical portfolio management process, people acquire public market information G together with
some private information Hin order to assess mand V. Considering the public information first, G typically
includes announced background economic driving information, historical market data, market consensus
(and maybe, mis-perception), and announced company-specific news etc. The information accrues over time
such that G
s
G
t
(time s < t). With only the common market information G, continue assuming normality,
the perceived security returns distribution can be represented by the estimated first and second moments,
2
In this paper, we use upper case R to stand for total return and lower case r for excess return, i.e., r = R r
f
.
3
In this paper, we use x to denote a random variable; x to denote a vector; and a bold symbol X to denote a matrix. The
dimension(s) of vector and matrix will be clarified on its first appearance. For example,

r
[n1]
stands for the n by 1 return vector with
random entries.
3
i.e.,

r
|G
N(

[n1]
,
[nn]
)
4
, where

= E(

r|G)
5
is the vector of mean estimates and = E(V|G) is the
variance-covariance matrix. The second-moment estimate is generally regarded as more reliable than the
first-moment estimates

. The latter is the holy grail of the investment industry.

On the other hand, the private information H generally includes particular insights of analysts which
are exclusively available to the PM. The insights usually come as a consequence of the analysts particu-
lar skills and efforts. Based on H, the PM forms her (private) view vector

y
[k1]|H,G
. These views are
then incorporated into the return forecast vector

m
[n1]
= E(

r|

y, H, G) with a revised covariance matrix

V
[nn]
= E(V|

y, H, G). Under normality assumption, these estimates fully characterise the distribution
of the security returns. Plugging

m and

V into a mean-variance optimiser, one solves a typical portfolio
optimisation problem.
In practice however, the incorporation of the public information G and the private information H in the
portfolio construction process is far from trivial. Many PMs focus on exploring public market data and that
acquired at a cost. Various quantitative techniques have been developed. Some commonly used extrapolation
techniques include, e.g., historical averages, equal means, risk-adjusted equal means
6
, or some modern time
series techniques with some prediction power etc.
Figure 1 shows the traditional process. Note views are formed drawing on various information sources,
and expressed in different forms, e.g., explicit returns, or relative performances or even in terms of strategies.
Before the BL model, it was not straightforward how to systematically convert such views or information
into explicit forecasts. Even in case of explicit return forecasts for some single security, there is still a lack
of mechanism to evaluate the implications of these signals for other securities induced by the dependence
structure.
Another, more fundamental, issue is that, the exploration of the commonly accessible market data might
be less productive than the private information. Economists argue that these techniques would not generate
insights superior to the market. In other words, if all we have are publicly available information and common
techniques, then why should we not just use the market view?
Black and Litterman (1992) (BL) take the point further and propose that without private views, the only
legitimate forecasts should be backed out from the market portfolio using the Capital Asset Pricing Model
(CAPM), the equilibrium pricer. In this case, it is optimal to simply use these forecasts to construct the
portfolio and manage it passively (by holding a slice of the market portfolio). With private information, the
forecasts should be updated based on the Bayes Rule, the fundamental law for belief updating. Therefore,
they recommend a Bayesian-analytic model for return forecasts and then resort to a conventional mean-
variance optimiser for portfolio construction. Consequently, the model addresses the lack-of-robustness
problem of portfolio optimisation through qualifying inputs rather than constraining the optimiser.
4
In this paper,

x
|I
is used to denote our perception of

x after examining the information I. Since in the Bayesian framework,
updated perceptions are still considered random, we need such notation to distinguish them from mean (point) estimates E(

x|I).
5
We use x to denote an estimate for x. Note in the traditional framework, estimates are considered deterministic; whereas in the
Bayesian framework, estimates are still random but with updated uncertainty.
6
See Black and Litterman (1992) for a description and critiques of these methods.
4
View generation
Portfolio optimisation
Public information
Portfolio
Optimiser
Private information
PM views
Explicit return forecasts
How to translate views into
explicit return forecasts?
zero forecast for no view?
Erratic outputs
Figure 1 The Traditional Approach to Portfolio Optimisation
Source: Nomura
3 Three Pillars: Semi-Strong Market Efficiency, CAPM, and Bayes
Rule
In the previous section, we refer to

r N( m, V) as the real distribution. Unlike in classical statistics
where the real means are considered deterministic (though unobservable), Bayesian statisticians consider
these as random themselves. Therefore, perceptions are always produced with some uncertainties, and are
represented by probabilities (i.e., probabilistic views). For example, after absorbing the public information
G, a PM forms her estimation

r
|G
N(

, ). As she is still not certain about this, particularly the mean

vector

, new information will always be sought after for further improvements.

Note the estimation

r
|G
N(

, ) only represents this investors personal perspective. She then uses the
estimation to interact with the market. In a marketplace with countless investors, there would be countless
estimates at a time. Suppose they disagree on the mean vector

, and so buy and sell into the market. The
market witnesses all these actions driven by individual predictions and prices these securities

r
M|

1
,

2
,...

N(

M
,
M
). The evaluation is beyond the ability of human beings and can only be undertaken by the
market.
Thanks to the CAPM, one can back-out the information from the market portfolio assuming equilibrium.
In other words, one may use the CAPM-assessed equilibrium estimation

r
e|G
N(

,
e
) as a proxy for
the market view

r
M|

1
,

2
,...
, where

is the vector of the CAPM-assessed excess returns
7
. Here, since the
market estimation is based on estimates of individual investors in the whole market, one should be more
7
Unless otherwise specified, in this paper, all returns refer to those excess of the risk-free rate
5
confident on

and thus the entries of

e
should be smaller than those of . It may therefore be argued that,
without any superior private insights, one should prefer

r
e|G
to her own estimation

r
|G
.
This basically means if all we have are public information and common techniques, the market may be
smarter than us. This argument reminds us of the semi-strong form of market efficiency hypothesis (Fama,
1970). The assumption suggests that all public information has been absorbed into the market pricing of
securities and can therefore not be explored to achieve abnormal returns. In other words, only with superior
private insight and techniques can the PM make superior returns.
In this vein, the BL assessment reduces to the CAPM-equilibrium layer in the absence of any private
information, where the security excess return vector can be backed-out from the market portfolio, i.e.,

[n1]
=

[n1]
[E(

R
M
|G) r
f
] (where

is the vector of security exposures to the market; E(

R
M
|G)
is the expected (full) market return; and r
f
is the risk-free interest rate). In the Bayesian framework,

is
considered as the prior estimate, since private information has not been examined yet.
After examining the private information H, PM forms a vector of views
8
,

y
|H,G
9
, usually in terms of
certain linear combinations of the securities in the universe, e.g., strategy, comparative performance views,
together with their performances. These have direct implications for the securities they cover and also, due
to the dependence structure, indirect implications for those in the universe not covered. So there is a need to
update her forecasts for the whole universe and the BL model rightly facilitates this. The updated forecasts
are denoted by

r
|

y,H,G
, and called the posterior estimates.
It should be noted that, in the BL model, the private information H is not directly fed into security
return forecasts

r
|H,G
; but is first absorbed into PM views

y
|H,G
, and then indirectly fed into the security
return forecasts. This is due to the belief that it is private and the direct security updates

r
|H,G
comes
only after private views, e.g.,

y
|H,G
, are digested by the market. However, the assumed view structure as
linear combination of security return forecasts is flexible enough to admit views in terms of explicit return
forecasts. So the linearity assumption is not restrictive.
The main contribution of the BL model is its unique insight enabling security returns to be assessed in the
Bayesian framework, drawing on some views on the underlying securities or their linear combinations. Its
closed-form solution for the posterior,

r
|

y,H,G
, serves as a smooth view blender enabling PMs to incorporate
their view vector

y
|H,G
into the portfolio construction and rebalancing process. (See Figure 2)
Comparative advantages of the model include: -
PMs no longer need to produce forecasts for the full universe of securities. Instead, providing any
number k (0 k n) of views suffices and these views can be relative (e.g., Security A will
outperform Security B by 2%) as well as absolute (e.g., Security A will grow by 10%).
The resultant allocations tend also to be more robust and intuitive. This is because the process starts
from a common CAPM-based equilibriumlayer. On top of this, tilts are generated to reflect the private
views.
8
Normally, analysts/PMs form their views according to information arrivals. In the remainder of this paper, we use view and
information interchangeably.
9
In the Bayesian framework, views should also be treated as random since there are always forecast errors.
6
View generation
Portfolio optimisation
Public information
Portfolio
Optimiser
Black-Litterman
Model
Systematic translation;
smooth view blending
Relatively stable
and intuitive outputs
Private information
Combined views
V

, ~
~
, , |
m N r
G H q
* *
*
PM views
) , ( ~
~
, |
q N y
G H
* *
Market view
(default view)
) , ( ~
~
W S P
* *
N
Figure 2 The BL-Based Portfolio Optimisation Framework
Source: Nomura
7
Now let us delve into the formulation of the model.
4 Model Formulation and Results
The model relies on two key technical assumptions.
Assumption 4.1 (Prior Return Forecasts) The prior return forecasts (based on the public information) are
normally distributed as follows:

r
|G

N(

, ) (1)
where is a positive multiplier applied to the estimated covariance matrix to proxy the prior error matrix.
Based on the economic reasoning in the previous section, this assumption is justified by: -
(a) With only public information G and supposing the market is already in equilibrium, the CAPM is
the legitimate market equilibrium pricer, and so

assesses

r.
(b) However, since the market is not necessarily in equilibrium, the assessment

suffers from errors

e
. Using a factor model-based argument, the model assumes
e
, the estimated security
covariance. Moreover, as explained in the previous section, the elements of
e
should be smaller
than those of in a market which demonstrates some level of semi-strong market efficiency; thus,
1
10
.
To assess distribution (1), we need to know the CAPM-based equilibrium return vector:

. Some linear
algebra drawing on the CAPM yields

= (
E(

R
M
|G)r
f

2
M
) w
M[n1]
. This allows us to back out security
returns from the current market portfolio w
M
. Based on

, PMs form the equilibrium layer of their portfolio

and then tilt their portfolio upon arrival of new private information.
Quite realistically, PMs make investment decisions relying on some (limited number of) theories, views
or strategies. Suppose a PM has k ( n) private views (or theories etc.) that are expressed or approximately
represented by some linear combinations of security returns:

P
[kn]

y
|

r[k1]
(2)
where

P is a matrix of view structure parameters or a vector of k strategies; and

y
|

r
represents the PMs
view forecast vector.
In (2), the views/theories can somehow be qualitative, and they need to be calibrated against available
information.
Upon arrival of the private information H, it is assumed that the PM still does not have explicit re-
turn estimates

r
|H,G
; yet (2) materialises to the extent that

P
|H,G
belief
P
[kn]
and thus, the updated view
becomes:
P

r
|H,G
=

y
|

r,H,G
+

[k1]
(3)
10
As we develop more insights in our forthcoming papers, we will see that is a subjective parameter and such restriction does not
exist.
8
where P is the concretised view structure matrix;

r
|H,G
is the (unknown but required) posterior vector of
return estimates;

y
|

r,H,G
is the vector of updated view estimation (given a realisation of

r); and

is the
vector of view estimation errors.
Note the structure of views as linear combination of prior return forecasts is flexible enough to accom-
modate relative as well as absolute views regarding security returns. For example, if the only view PM has
is that, with a standard deviation of 1%, security A will outperform security B by 2%. This relative view can
be expressed as:
_
1 1
_
_
_
E(

r
1
|H, G)
E(

r
2
|H, G)
_
_
= 2% + where N(0, (1%)
2
)
Alternatively, P can also be viewed as k strategies formulated based on the new information. In this
example, 2% can be considered as the expected return of the spread strategy
_
_
1
1
_
_
.
For the view estimation error vector

in (3), the model makes the following normality assumption:

Assumption 4.2 (View Errors) The view error vector is normally distributed as follows:

[k1]
N(

0
[k1]
,
[kk]
) (4)
where

0 is a vector of zeros; and is a diagonal variance matrix of view-estimation errors, which are
considered, for simplicity, as independent across views
11
.
Therefore, we have

y
|

r,H,G
N(P

r
|H,G
, ) (conditional on

r). With our prior knowledge (1), plus
these views and our final conviction

y
|

r,H,G
belief

q
[k1]
(the ultimate view mean estimation), the Bayes
Rule can be utilised to leverage (3) for an update of the forecasts of

r
12
. The closed-form results are:
Theorem 4.1 (Posterior Return Estimates) The posterior return vector is also normally distributed, i.e.,

r
|

q,H,G
N(

m,

V), where the updated mean vector is:

m =
_
()
1
+P
T

1
P

1
_
()
1

+P
T

q
_
(5)
and the updated variance-covariance matrix is:

V =
_
()
1
+P
T

1
P

1
(6)
Proof. See Appendix A.
11
This assumption comes as a convention of the classical linear regression model. This provides some simplicity. However, as far as
the model is concerned, this independence restriction is not necessary.
12
Careful readers may have noted the relationship (3) can be re-arranged and considered as a linear system

y
|

r,H,G
= P

r
|H,G

for fitting

r
|H,G
, with

q and P as data. It seems that

r
|H,G
can be estimated by the generalised least square (GLS) estimator

P
T
[nk]

1
[kk]
P
[kn]

1
P
T
[nk]

1
[kk]

q
[k1]
. There are however two reasons why the following BL solution is more practical:
(1) In practice, it is often the case that we only have k (n) views. As such, the matrix P
T

1
P is singular and therefore the GLS
estimation is not attainable. (2) Treating

q and Pas data is too optimistic - these are often just analysts opinions. The strong uncertainty
associated makes the Bayesian technique a more natural fit.
9
The results are intuitive. Consider the 1-security, 1-view scenario:
m =

2
+
q

2
1

2
+
1

2
=
1
+
2
q (7)
or
posterior estimate = confidence-weighted average of views; and
1

V
=
1

2
+
1

2
(8)
or
posterior confidence = aggregate of confidence,
where is the security volatility; and is the view estimation error, and
1
=
1

2
1

2
+
1

2
and
2
=
1

2
1

2
+
1

2
are the confidence weights.
If we consider confidence as the inverse of variance (the uncertainty), Equation (8) basically blends
the PMs confidences on the market equilibrium view,
1

2
, and her own view,
1

2
, to obtain the posterior
estimation confidence as
1

2
+
1

2
. This simple additive relationship comes as a result of the Bayess Rule.
The more confident the PM is on either view, the more confident she will be on her posterior (8).
Equation (7) simply combines the market equilibrium with the PM views through a confidence-weighted
average scheme. Again, this relationship is dictated by the Bayes Rule. It is easy to see that the more
confident the PM is about her view q, the more weight she should put on the view and therefore her posterior
forecasts should be adjusted more towards q. Consequently, the more she should tilt her new portfolio
towards reflecting the view; otherwise, the more she should rely on the market portfolio. In the extreme case
where her confidence on the views is minimal (or she has no views), Theorem 4.1 reduces to (1) and she
should do nothing to the equilibrium layer.
5 Implementation of the Black-Litterman Model
The BL framework in practice involves 4 steps:
Step 1: Data collection for the market-wide variables, portfolio-specific variables, and the PM views.
The following inputs are needed:
General economy-related input:
r
f
: the risk-free interest rate
Benchmark portfolio-related inputs: need to pick an index type of portfolio representing the
universe and the market portfolio, with which, the following are obtained:
w
M
: the market portfolio weights
E(

R
M
|G): the estimated total market portfolio return
10
Portfolio-related inputs:
: the risk multiplier. The value choice for this parameter should be considered
together with the view-uncertainty matrix to achieve a desired balance for
shinkage
: the estimated security return variance-covariance matrix
View-related inputs:
P: the view/strategy structure

q: the view estimates, i.e., estimated security/strategy returns

: the diagonal matrix containing the view variances
Step 2: Back-out the market view from the market portfolio based on the CAPM.
This basically requires the assessment of the CAPM-based excess returns

=
E(

R
M
|G)r
f
w
T
M
w
M
w
M
.
Step 3: Update the forecasts according to Theorem 4.1.
This involves substituting the results from Step 2, together with other inputs as specified in Step
1, into Theorem 4.1 to evaluate the posterior mean

m and error matrix

V.
Step 4: Optimise the allocation based on the posterior estimates to decide how to tilt from the market
portfolio.
The general mean-variance optimisation problem is:
max
w
{ w
T

m w
T

V w} (9)
Since most of the data items can be obtained from a market database, the PM just needs to pick an
appropriate risk-aversion parameter and concentrate on the view-related inputs: , P,

q and .
Figure 3 illustrates how the user should prepare inputs and how the model interacts with other compo-
nents in a typical equity analytical framework.
6 Practical Issue: Curse of Dimensionality
Note the posterior covariance matrix

V
[nn]
produced by the BL model is a fully populated numerical
matrix. When n > 1000, this heavy matrix poses significant computational challenge to any quadratic
optimiser. In order to leverage the capacity of the optimiser, dimension reduction may be needed.
We recommend the following eigensystem analysis. Since

V is real, symmetric and positive-definite,
the following diagonalisation is guaranteed:

V = E
[nn]
D
[nn]
E
T
[nn]
(10)
where Dis the diagonal matrix with the eigenvalues sorted from high to low; and E represents the loadings
matrix constructed by the eigenvectors corresponding to the sorted eigenvalues.
11
Black-Litterman Model
M M M
w w
* *

T
= o ( )
M
r
w
M
G M
*
*

2
|

o
t =
M G M
r o
2
1
|
~
] ) [( ] ) [(
1 1 1 1 1
q m
T T
* * *

+ + = P P P t t t
1 1 1
] ) [(

+ = P P V
T
t
Optimised portfolio weight
Dimension high?
Benchmark
and Risk
Model
Database
User
Benchmark?
Index
Choose from a list
Portfolio:
Copy & paste
Identifiers
Linear PM Views
Risk model choice
] [ n k
P
] 1 [

k
q
*
] [ k k

] [ n n

) , ( ~
~
, |
P q N r
G H
* *
Portfolio
Optimiser
Dimension Reduction for
m
*

Yes
V

] [
T
] [ ] [ ] [

n n n p p p p n
+ ~ V
m
*


No
m
*
V

] 1 [ n M
w
*
n informatio public :
return excess portfolio market :
operator n expectatio : ) E(
forecasts market of vector :
|
G
r
G M

t
*
scalar confidence w market vie :
weights portfolio market :
covariance - ance stock vari :
risk market :
t
o
M
M
w
*

matrix error posterior of matrix :

estimates view :
matrix y uncertaint view :
matrix structure view :
estimates mean posterior of vector :
V

P
q
m
*
*
1
default
= t
*
w
*
Figure 3 Interactions between Different Components in the BL Framework
] 1 [ n M
w
*
12
Using the first p (p n) eigen components to approximate the original matrix and collect residuals, we
have:

V =
[np]

[pp]

T
[pn]
+R
[nn]
(11)
where is the diagonal matrix reduced fromDby retaining the first p engenvalues; is the loading matrix
reduced from E by retaining the first p eigenvectors; and Rrepresents the residual matrix.
Impose the following treatment:

V
T
+
[nn]
(12)
where is the diagonal matrix obtained from Rby setting all its off-diagonal elements to 0.
By such approximation, we attain a sparse representation of

V (i.e., is just a [n p] matrix, is a
small [p p] matrix, and is a diagonal matrix). This will significantly help the optimiser.
With the same choice of p, the approximation has exactly the same quality as a principal component
analysis (PCA) risk model. In risk modelling, the application of the eigensystem analysis is meant for noise
disposal. So generally, a very limited number l of principal components enter in
T
. In our case,
however, the main purpose is to reduce the dimension to the level that the optimiser can handle. Therefore,
the choice of p is only subject to the optimiser capacity, which can easily handle several hundreds. We may
therefore set p to a number much larger than l (e.g., p = 100) to allow a very good approximation of the
posterior covariance matrix.
Alternatively, to avoid the dimension issue, one may also rank the securities according to the posterior
estimates and reduce the universe before optimisation.
7 Conclusion
This paper is an introduction to the Black-Litterman (BL) Model. Essential technical details, i.e., motivation,
intellectual roots, model formulation, and implementation guidance, are included with a view to enhancing
appreciation and enabling implementation of the model.
In addressing the lack-of-robustness issue of the mean-variance optimiser, the BL model aims to qualify
the view inputs rather than constraining the optimiser. In a market that demonstrates semi-strong market
efficiency, the market is smarter than any individual with only public information and conventional tech-
nique; therefore, one should simply rely on the market view, implied from the market portfolio by the
CAPM, the equilibrium pricer. Only when an investor has unique insight and superior forecasting technique
can she potentially outperform the market. To take advantage of these, the Bayes Rule, the fundamental Law
of belief updating, is employed for view processing (i.e., updating, translation, blending and shrinkage). It
is based on these three notions that the BL model is established.
Drawing on normality and linearity assumptions, this model admits analytical solution. Implementation
can be efficient. Since the outputs are in the form of explicit return forecasts together with a covariance
matrix, a standard mean-variance optimiser can produce allocation recommendation which is considered
13
robust due to the shrinkage effect. This emancipates PMs from the job of view processing and portfolio
construction, enabling them to concentrate on alpha generation.
However, the user should be aware of some BL practical issues.
Curse of dimensionality. The BL model spits out a fully populated numerical posterior matrix. In case
of a large universe, this poses significant burden to the optimiser. We resolve this issue by obtaining a
sparse representation of the matrix drawing on the principal component approximation.
Confidence parameter setting. Setting and is a classical issue inherited from the Bayesian frame-
work.
View correlations. For simplicity, the BL model treats views as orthogonal. In reality, view correla-
tions may exist but hard to quantify. Is this a problem?
Prior setting. In practice, there are a variety of investment styles. Equilibrium is an abstract concept;
and the market portfolio does not seem to be the starting point suitable for all strategies. How should
we choose an appropriate prior?
Factor-based portfolio construction. How can we apply the BL technique to the popular Fama-French
factor ranking approach?
Optimiser issues. The BL framework calls an optimiser for allocation purpose. Whereas the opti-
miser behaves as a black box. There are risk-aversion as well as intuition issues. Can we make the
framework more open and intuitive?
Risk model quality. Risk model has errors itself. How can we economically use it in allocation?
Linearity and normality assumptions. In the BL model, security returns are considered normal; and
the factor model, linear. How to apply this model to non-normal, non-linear markets?
In our forthcoming pieces, we aim to make the BL technique practical by resolving all these issues.
14
References
[1] Black, F. and R. Litterman (1992), Global Portfolio Optimisation in Financial Analysts Journal,
Sep./Oct. 1992, Vol.48, No.5, pp. 28-43
[2] Cheung, W. (2009a), The Black-Litterman Model Explained (I), Nomura 2009 (first published in
2007)
[3] Cheung, W. (2009b), The Black-Litterman Model Explained (II), Nomura 2009 (first published in
2007)
[4] Fama, E. F. (1970), Efficient Capital Markets: A Review of Theory and Empirical Work in Journal
of Finance, Vol.25, No.2, pp.383-417
[5] Hamilton, J. (1994), Time Series Analysis, Princeton University Press: Princeton, NJ
[6] He, G. and R. Litterman (2003), The Intuition Behind Black-Litterman Model Portfolios Available
at SSRN: http://ssrn.com/abstract=334304 or DOI: 10.2139/ssrn.334304
[7] Idzorek, T. M. (2004), A Step-by-Step Guide to The Black-Litterman Model, Zephyr Working Doc-
ument 2004
[8] Jones, R., T. Lim and P. J. Zangari (2007), The Black-Litterman Model for Structured Equity Portfo-
lios in Journal of Portfolio Management, Winter 2007, pp.24-33
[9] Lee, W. (2000), Advanced Theory and Methodology of Tactic Asset Allocation, forth-coming book
[10] Martellini L. and V. Ziemann (2007) Extending Black-Litterman Analysis Beyond the Mean-Variance
Framework, Working Document, EDHEC
[11] Meucci, A. (2005), Risk and Asset Allocation, Springer: New York
[12] Meucci, A. (2006), Beyond Black-Litterman: Viewon Non-normal Market, in Risk, Feb 2006, pp.87-
92
[13] Satchell, S. and A. Scowcroft (2000), A Demystification of the Black-Litterman Model: Managing
Quantitative and Traditional Portfolio Construction, in Journal of Asset Management, Vol.1, No.2,
pp.138-50
[14] Zhou, G. (2008), An Extension of the Black-Litterman Model: Letting the Data Speak, Working
Document, Washington University in St. Louis
15
Appendix A: Proof
Theorem 4.1 (Posterior Return Estimates) The posterior return vector is also normally distributed, i.e.,

r
|

q,H,G
N(

m,

V), where the updated mean vector is:

m =
_
()
1
+P
T

1
P

1
_
()
1

+P
T

q
_
(13)
and the updated variance-covariance matrix is:

V =
_
()
1
+P
T

1
P

1
(14)
Proof. We first deal with a general case and then apply it into the BL context. In general, assume the
following linear regression relationship:

Y
[k1]
= Q
[kn]

X
[n1]

u
[k1]
(15)
where Q is a known/observable matrix;

Y is an observable vector;

X is unobservable and needs to be

estimated; and

u is the error vector.

Also, assume the following Gaussian distributions for regression errors and the prior:

u N(

0
[k1]
,
[kk]
); and (16)

X N(
X[n1]
,
X[nn]
) (17)
We therefore have, conditional on a realisation of

X,

Y
|

X
= Q

u N(Q

X, ) (18)
Distributions (17) and (18) are equivalent to:
f(

X) =
1
(2)
n
2
|
X
|

1
2
e

1
2
(

X
X
)
T
(
X
)
1
(

X
X
)
(19)
where | | gives the determinant of the matrix it applies to; and
f(

Y |

X) =
1
(2)
k
2
||

1
2
e

1
2
(

Y Q

X)
T
()
1
(

Y Q

X)
, (20)
respectively.
We try to imply the probability distribution of

X
|

Y
from the joint probability density function:
f(

X,

Y ) = f(

X)f(

Y |

X)
=
1
(2)
n+k
2
|V
a
|

1
2
e

1
2

T
V
1
a

(21)
where

[(n+k)1]
=
_
_

X
X

Y Q

X
_
_
; and (22)
16
V
a[(n+k)(n+k)]
=
_
_

X
0
[nk]
0
[kn]

[kk]
_
_
(23)
We hope, and based on the Bayes Rule, it is possible to express (21) in terms of f(

X|

Y )f(

Y ) since we
are interested in the posterior estimation of

X given

Y . This translates into a need to replace the dependence

of the mean estimates of

Y on

X (as in (18)) with a dependence of the mean estimates of

X on

Y . In other
words, through some transformation, we need to get rid of

X from the second error vector, but allow

Y to
enter the first error vector. To this end, construct the following matrix (Hamilton, 1994, Ch.12), which can
help us achieve this:
A
[(n+k)(n+k)]
=
_
_
I
[nn]
[(
X
)
1
+Q
T

1
Q]
1
Q
T

1
0
[kn]
I
[kk]
_
_
_
_
I
[nn]
0
[nk]
Q
[kn]
I
[kk]
_
_
=
_
_
[(
X
)
1
+Q
T

1
Q]
1
(
X
)
1
[(
X
)
1
+Q
T

1
Q]
1
Q
T

1
Q
[kn]
I
[kk]
_
_
(24)
From the first equality, we may check |A| = 1 and it can be shown that the transformed error vector
becomes:

= A
=
_
_

X [(
X
)
1
+Q
T

1
Q]
1
[(
X
)
1

X
+Q
T

Y ]

Y Q
X
_
_
=
_
_

X

m(Q,
X
, ,
X
,

Y )

Y Q
X
_
_
(25)
where

m(Q,
X
, ,
X
,

Y ) = [(
X
)
1
+Q
T

1
Q]
1
[(
X
)
1

X
+Q
T

Y ];
and the transformed error covariance matrix becomes:
V

a
= AV
a
A
T
=
_
_
[(
X
)
1
+Q
T

1
Q]
1
0
[nk]
0
[kn]
Q
T

X
Q+
_
_
=
_
_

V(Q,
X
, ) 0
[nk]
0
[kn]
Q
T

X
Q+
_
_
(26)
where

V(Q,
X
, ) = [(
X
)
1
+Q
T

1
Q]
1
.
17
Through these transforms and noting |A
1
| = |A| = 1, (21) can be rearranged as follows:
f(

X,

Y )
=
1
(2)
n+k
2
|V
a
|

1
2
e

1
2

T
V
1
a

=
1
(2)
n+k
2
|(A
1
)V

a
(A
1
)
T
|

1
2
e

1
2
(A
1

)
T
[(A
1
)V

a
(A
1
)
T
]
1
(A
1

)
=
1
(2)
n+k
2
|V

a
|

1
2
e

1
2
(

)
T
(V

a
)
1
(

)
(27)
f(

X,

Y )
=
1
(2)
n+k
2

V() 0
[nk]
0
[kn]
Q
T

X
Q+

1
2

1
2

X

m()

Y Q
X

V
1
() 0
[nk]
0
[kn]
(Q
T

X
Q+)
1

X

m()

Y Q
X

=
1
(2)
n
2

V()

1
2
e

1
2

m()

V
1
()

m()

1
(2)
k
2

Q
T

X
Q+

1
2
e

1
2

Y Q
X

T
(Q
T

X
Q+)
1

Y Q
X

(28)
where

V() =

V(Q,
X
, ); and

m() =

m(Q,
X
, ,
X
,

Y ).
From (28), we see:

X
|

Y
N
_

m(Q,
X
, ,
X
,

Y ),

V(Q,
X
, )
_
(29)
and

Y N
_
Q
X
, Q
T

X
Q+
_
(30)
To assess

X
|

Y
in (29), people use their best knowledge regarding Q,
X
, ,
X
, and

Y .
Recall in the model setting, our best knowledge based on the public information G leads to the following
prior belief:

r
|G
N(

, ) (31)
After examining the private information H, we form the following updated views:
P

r
|H,G
=

y
|

r,H,G
+

(32)
where P is the view structure;

y
|

r,H,G
is the view forecast vector; and suppose

N(

0, ).
Therefore, conditional on a realisation of

r
|H,G
:

y
|

r,H,G
= P

r
|H,G

N(P

r
|H,G
, ) (33)
18
Substituting

r for

X and

y for

Y into (29), we reach:

r
|

y,H,G
N
_

m(P, , ,

y
|

r,H,G
),

V(P, , )
_
= N
_
[()
1
+P
T

1
P]
1
[()
1

+P
T

y
|

r,H,G
], [()
1
+P
T

1
P]
1
_
(34)
Finally, using our eventual conviction about the mean of

y
|

r,H,G
belief

q, we reach the following posterior
belief:

r
|

q,H,G
N
_
[()
1
+P
T

1
P]
1
[()
1

+P
T

q], [()
1
+P
T

1
P]
1
_
(35)
This completes the proof.
19